General Motion Classifier

Introduction

General Motion Classifier is software that classifies repetitive motions in real-time over BLE communication with the STMicorelectronics Sensortile, using optional user-defined Keras + Scipy models.

The required libraries and packages are: Keras (Python), LiquidDSP (C), gatttool, bluetoothctl, and expect. The software also requires the Sensortile to be running code found in UCLA STMicroelectronics Tutorial 8. Please refer to the following link for this: http://iot.ee.ucla.edu/UCLA-STMicro/index.php/UCLA_-_STMicroelectronics .

The program allows the user to define their own machine learning and signal processing classification models in the Python (See Custom Models). But, by default, the current program comes with the LSTM and the CNN+LSTM machine learning classification models.

Note, models do not interact with raw sensor data. Rather, raw sensor values are first passed through a low- pass filter and optionally bias-corrected (See Flags and Extensions). Furthermore, as of now, the program only uses x, y, and z acceleration data from the sensor, discarding the other sensor values.

Note, the current program cannot run on the Beaglebone platform as the space required to store Keras is too large (despite Keras/Theano supporting 32bit ARM systems). Rather, it is meant for prototyping valid machine learning models before being ported onto embedded platforms. The current program also cannot run on Windows systems due to the use of POSIX-defined multi-processing, synchronization, and IPC.

Dependencies

Keras v2.0.8
Tensorflow (Keras backend) v1.8.0
LiquidDSP v1.3.1

Getting Started

To run General Motion Classifier with the default CNN+LSTM model, first modify bctl.txt such that the first line is the MAC address of your computer and the second line is the MAC address of your Sensortile:

Next, run the following command: $ make program

Turn on your Sensortile using software from Tutorial 8 of the UCLA STMicroelectronics Tutorials.

Run the following command: $ ./main (or $ sudo ./main)

Flags and Extensions

./main [-r] [--bias=[status]]

The -r flag will skip the training step and perform real-time classification with an existing model.m5 file. Note, the user will still need to enter the names of the motions for the program to match the model predictions with human-readable mnemonics.

The --bias=off flag will turn off bias-correction in the DSP driver (See Program Flow).

The --bias=save flag will save the bias weights used by the DSP driver in a bias.config file.

The --bias=load flag will load saved bias weights from a bias.config file.

Acknowledgements

Thanks to Professor Kaiser for guidance with project.
Credit to Charles Zaloom for initial LiquidDSP code.